GenericDescriptorMatcher for Fern algorithm

edit

Thank you for your interest.

It's a summary of my code:

vector<cv::Mat> images; vector< vector<cv::KeyPoint> > keypoints;

descriptorMatcher = cv::GenericDescriptorMatcher::create("FERN", filename); descriptorMatcher->add(images, keypoints); descriptorMatcher->train();

cv::Mat testImage = cv::imread("test.jpg"); vector<cv::KeyPoint> testKeypoints;

descriptorMatcher->classify(testImage, testKeypoints);"

The keypoints are found with SURF detection.

In the "classify" method, it will have to be the query image, right? So why there are an error with this variable (images) ?

I have not enough experience with this code, but what i find weird is you training a matcher with a vector of keypoint vectors. Could it be that the inner vector should be a Mat element with features detected on a row element?

@fuego: Interesting, haven't used Fern so far either (maybe because they are under legacy). I guess you just abbreviated your code but for being sure: The vector of images contains your images and the vector of keypoints contains the computed keypoints for these images? Because the error may mean that your keypoints don't fit to your images (the same may be for you call to 'classify': Does the keypoints belong to this image?)

To answer your question in your comment: yes, according to the documentation you should pass the query image for 'classify'.

@steven: I don't think that you give features to the 'fern'-classifier, according to http://www.computer.org/csdl/trans/tp/2010/03/ttp2010030448.pdf which is referenced in modules/legacy/src/planardetect.cpp it classifies the patches surrounding the keypoints which you need to pass to the Fern-classifier.

Quanta : Exactly, the images are the train images with their respective keypoints (for each one). And "testKeypoints" are the computed keypoints of "testImage". I have checked that the keypoints are not out of the images.

So i don't know why this error appears.

Adding more precision to my application, i try to find an object in a panorama. However the panorama it is not just one image but many tiles. Consequently, my test image is not exactly the same of a tile but it looks like one of them.

Any others suggestions?

Sry, currently not, however I am very curious about it myself. Maybe you want to post the question to stackoverflow? Maybe I'll also have the time to try it out myself during the next week, but I cannot promise it - much to do :(